Simulations of gas pipelines and road traffic run conservation laws, which are equations for how density and momentum flow along a line, on networks of one-dimensional segments. The weak point of these simulations is not the pipes but the places where segments meet: junctions, and obstacles such as a slow bus that move along a segment. At every such interface the scheme has to solve an algebraic "coupling condition" that says how the states on each side fit together. Current schemes solve it once per step from frozen, piecewise-constant data, so they are only first-order accurate in time at the node, even when the pipe interiors are second order. Two recent network schemes state this limitation explicitly: A central scheme for coupled hyperbolic systems and An asymptotic-preserving scheme for isentropic flow in pipe networks. The obvious fix is a higher-order Runge–Kutta method, but that means solving the nonlinear coupling at every internal stage. Stages are exactly where boundary layers and spurious peaks appear.
We propose to borrow a lesson from geometric numerical integration, the study of time-steppers that preserve physical structure such as energy. That field has proved a sharp trade-off. The cheapest time-stepper that conserves energy exactly, the averaged vector field (AVF) method, is unique and cannot exceed second order (Sun et al.). Symplectic mono-implicit Runge–Kutta methods are likewise capped at order two (Noren et al.). Our claim is that at network interfaces this ceiling is the right target. A single AVF-type step that imposes the coupling condition once, on time-averaged interface traces, should give second order in time at an equal-pressure junction. The same step should reproduce the junction's energy balance exactly at the discrete level. We then use moving bottlenecks as a stress test. There the interface is non-smooth, because a flux constraint switches on and off, and the order barrier bites hardest.
Core Claim
Consider isentropic gas flow in pipes joined at a node. Each pipe has density $\rho_i$, velocity $v_i$, mass flux $q_i=\rho_i v_i$ and enthalpy $h(\rho_i)$. The equal-pressure coupling condition requires all pipes to share one pressure at the node. Together with mass conservation $\sum_i q_i = 0$ (signed by orientation), this forces equal enthalpies. The net energy flux into the node is then
$$ \Delta E_{\text{node}} = \sum_i q_i \frac{v_i^2}{2}, $$
which is generally nonzero and can take either sign. This quantity is the physical energy defect of the equal-pressure closure. Banda et al. showed that this closure prevents the non-physical pressure jumps allowed by a subsonic-flow closure, but it does not conserve energy. A trustworthy scheme should reproduce $\Delta E_{\text{node}}$ and add no numerical energy production of its own.
Our time step has the AVF form
$$ \frac{u^{n+1}-u^n}{\Delta t} = \int_0^1 F\big(u^n + \tau(u^{n+1}-u^n)\big)\,d\tau . $$
Here $u$ is the vector of cell averages on all pipes and $F$ is the semi-discrete right-hand side, which includes the node fluxes. The coupling condition is imposed on the traces of this averaged state, not on stage values. If $F$ is written as a skew part plus a dissipative part acting on the gradient of the discrete energy, the step gives an exact discrete energy identity. Interior dissipation is non-positive, and the node term is the discrete analogue of $\Delta E_{\text{node}}$.
The step costs one nonlinear network solve, is second order, and makes the node's energy defect a checkable diagnostic. It differs from the two network schemes above, which are first order at the node and never examine energy. It also differs from Noren et al., who study the order–structure trade-off only for lumped ordinary differential equations without interfaces.
Research Program
The first task is to build a semi-discretization of isentropic Euler on a network in the skew-plus-dissipative gradient form the AVF step needs. The interior uses entropy-conservative interface fluxes plus explicit dissipation. The node flux is derived from the equal-pressure half-Riemann problem, so that the energy identity at the node is inherited rather than imposed.
For isothermal pressure the energy contains $\rho\log\rho$, a transcendental term. Celledoni et al. established that finite-stage exactness holds only for polynomial Hamiltonians. We therefore evaluate the averaging integral by Gauss quadrature and measure the resulting energy error rather than assume it away.
We will prove second-order consistency of the averaged coupling for smooth traces and a discrete energy inequality in general. We will also test empirically whether the fully implicit step keeps its accuracy in the low-Mach limit, where Redle & Herty need a dedicated asymptotic-preserving splitting. We do not assume the AVF step is asymptotic-preserving. If it is not, we fall back to AVF on the slow part inside their splitting.
The second task transfers the same step to the coupled PDE–ODE moving-bottleneck model: Lighthill–Whitham–Richards traffic plus a vehicle trajectory with the unilateral flux constraint. Across the resulting non-classical shock, averaging a vector field that switches between constrained and unconstrained branches breaks the smoothness that the AVF theory assumes. We handle this by freezing the set of active constraints over a step and locating switching times by root-finding on the constraint residual. The step is then restarted at the event, and within each smooth sub-interval the averaged coupling is applied.
The in-cell reconstruction of Chalons et al. is kept unchanged for the spatial jump. What changes is the time coupling between the partial differential equation and the ordinary differential equation. The hypothesis is that the "small spurious oscillations" at constraint activation reported by Goatin et al. come from their explicit Euler trajectory update being out of step with the flux constraint. Under that hypothesis, event-aligned averaged coupling should remove them.
Bottlenecks crossing junctions are a separate problem. A cell then contains two discontinuities, and the question of whether a solution exists is still open (Nick Zinat Matin & Delle Monache). We treat that case only as an exploratory final run.
Baselines
The naive reference is the first-order node coupling of Herty et al.. We use their consistent nonlinear relaxation coupling ("Approach 4"), which is reported to remove interface artifacts and to converge at first order in $L^1$. No reported result was found on Lacuna for its absolute errors or run time, so we regenerate those numbers ourselves.
The strongest existing competitor on stiff networks is the asymptotic-preserving implicit-explicit (IMEX) scheme of Redle & Herty. They report first order in time and second order in space away from junctions, and about 140 times faster than an explicit scheme at $\epsilon=0.001$ on T-junctions. We run their scheme and ours on the same 1-to-2 and 2-to-1 T-junction configurations, with the same spatial reconstruction and grids. We compare at matched final-time error, not matched step size. Beating it means one of two outcomes. The first is reaching a target node-trace error with fewer total flops. The second is equal cost with an observed temporal order near two at the node. Parity in the stiff regime would already be a win if node order doubles.
The ablation baseline is the most dangerous one. It is a standard second-order strong-stability-preserving Runge–Kutta method, and a fourth-order Runge–Kutta method, each solving the equal-pressure coupling at every stage with the same nonlinear solver. This isolates the key ingredient: coupling once on averaged traces versus coupling per stage.
A second ablation replaces AVF with the implicit midpoint rule. Midpoint is symplectic and second order but conserves only quadratic invariants, and the gas energy is not quadratic. This separates the effect of "implicit and symmetric" from that of "energy-exact."
For moving bottlenecks, the baselines are two schemes. One is the in-cell reconstruction with forward trajectory integration of Chalons et al., which reports $L^1$ convergence between $O(\Delta x^{0.8})$ and $O(\Delta x^{1.5})$, sharp jumps and exact mass conservation. The other is the explicit-Euler constrained Riemann scheme of Goatin et al.. Beating them means keeping their sharpness and mass conservation while reducing overshoot at activation and deactivation, at equal grids and equal CFL number.
Why not just…?
Why not just use a standard high-order Runge–Kutta method with the coupling solved at every stage? This is the real threat, and it is the per-stage ablation above. The evidence against it is indirect. Herty et al. show that inconsistent interface closures create boundary layers and peaks, and stage values are not physical states at the stage time. The order-barrier results suggest that any extra accuracy is unlikely to come with structure. If per-stage second-order Runge–Kutta reaches second order at the node without artifacts and at lower cost, the order part of our claim falls. Only the energy-diagnostic claim would then remain, and we would report that outcome.
Why not just take smaller first-order steps? Under acoustic stiffness, explicit steps are CFL-limited, which is the 140-fold gap reported by Redle & Herty. Implicit first-order steps pay error linearly in $\Delta t$. The matched-error cost curves in our experiments settle this directly. We expect a crossover only at loose tolerances.
Why not a subsonic-flow or linear relaxation closure, which is easier to average? Banda et al. show that the subsonic closure admits non-physical pressure jumps. Herty et al. show that linear coupling on relaxation variables produces artifacts. Cheapness at the node is not worth losing consistency.
Why not go to fourth order, as the Mean Inverse Integrator does for noisy data? Noren et al. find that fourth order beats structure preservation for filtering noise in smooth ordinary differential equations. Interfaces are not smooth, however. Our fourth-order Runge–Kutta ablation tests whether that accuracy survives at the node. We predict it drops to between one and two, which would make order two with exact energy the better trade.
Proposed Experiments
On T-junction networks with smooth inlet data, we measure temporal convergence of node traces against a fine reference. The decisive check is an observed order of at least 1.8 for averaged coupling, against about one for the Redle–Herty scheme. We also compare that order with the order obtained from per-stage second- and fourth-order Runge–Kutta.
On each run we monitor the discrete node energy residual. This is the discrete energy change minus interior dissipation minus the discrete $\Delta E_{\text{node}}$. Averaged coupling should hold it at quadrature or solver tolerance, while the baselines show an $O(\Delta t)$ spurious term. A sign test on flows where $\Delta E_{\text{node}}>0$ checks whether the baselines create energy beyond the physical defect.
We sweep $\epsilon$ down to 0.001 and plot cost against error to test stiffness robustness.
For moving bottlenecks, we rerun the five Riemann cases used by Chalons et al.. We add the multi-vehicle activation and deactivation scenarios of Goatin et al., measuring maximum overshoot and total-variation increase near switching events. We also measure $L^1$ convergence and mass error. The decisive check is that event-aligned averaged coupling removes the reported oscillations without reducing the convergence range. If the oscillations persist, their origin lies in the spatial reconstruction rather than the time coupling. That would itself be a useful negative result for the open stability question.
References
- A Central Scheme for Two Coupled Hyperbolic Systems
- An Asymptotic-Preserving Scheme for Isentropic Flow in Pipe Networks
- The minimal stage, energy preserving Runge-Kutta method for polynomial Hamiltonian systems is the averaged vector field method
- Learning dynamical systems from noisy data with inverse-explicit integrators
- Coupling conditions for gas networks governed by the isothermal Euler equations
- Energy-preserving Runge-Kutta methods
- A conservative scheme for non-classical solutions to a strongly coupled PDE-ODE problem
- Interacting moving bottlenecks in traffic flow
- On the Existence of Solution of Conservation Law with Moving Bottleneck and Discontinuity in FLux